TBM remote monitoring method based on multi-level collaborative management architecture

Abstract

The present application relates to the technical field of remote monitoring, and more particularly to a TBM remote monitoring method based on a multi-level collaborative management architecture, applied to a remote comprehensive monitoring platform. The platform comprises a construction operation layer, a construction management and control layer, and a decision support layer. Each monitoring scene is composed of multiple construction surface nodes to form a local blockchain ring, and the whole forms an overall blockchain ring. The method comprises: collecting construction state data and standardizing processing, inputting into a device risk identification model and a geological evaluation model for construction state analysis; based on local anomaly detection, deducing an abnormal propagation path, calculating the reverse coverage capability of each non-anomalous local ring, and determining a decision blockchain ring; through abnormal isolation control intelligent contract, node locking and decision switching are carried out, and a decision instruction based on a multi-objective optimization algorithm is generated. The present application improves the abnormal self-healing ability and dynamic decision efficiency of the TBM construction monitoring system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of TBM remote monitoring, and particularly relates to a TBM remote monitoring method based on a multi-level collaborative management architecture. BACKGROUND

[0002] As a large-scale tunnel construction equipment integrating mechanical, electronic, hydraulic, sensing and other technologies, TBM (Tunnel Boring Machine) is mainly suitable for tunnel excavation of hard rock and extremely hard rock. In TBM tunnel construction, there are multiple construction operation faces including tunneling, supporting and slag removal, etc., so it is necessary to remotely monitor TBM construction to coordinate construction progress, monitor equipment safety and ensure construction quality. Existing TBM remote monitoring systems mostly adopt centralized or single-level monitoring architecture, which collects equipment operation parameters, geological environment parameters and construction efficiency parameters collected by each monitoring point through a central server, and performs abnormal detection and instruction issuing based on a preset risk model or rule. However, in complex construction environments such as super-long tunnel excavation and highland geological construction, the construction state is highly heterogeneous and changes frequently, and the single-level centralized processing method faces problems such as rapid expansion of data volume, heavy load of central processing, increased response delay and inability to isolate abnormality in time, which seriously restricts the stability and decision efficiency of the remote monitoring platform. SUMMARY

[0003] The present application provides a TBM remote monitoring method based on a multi-level collaborative management architecture to solve the technical problems of the prior art. The method is applied to a remote comprehensive monitoring platform. The platform includes a construction operation layer, a construction control layer and a decision support layer. Each monitoring scene is composed of a local blockchain circle by multiple construction face nodes, and the whole forms an overall blockchain circle. The method includes: collecting construction state data and standardizing processing, inputting into an equipment risk identification model and a geological assessment model for construction state analysis; based on local anomaly detection, deducing an abnormal propagation path, calculating the reverse coverage ability of each non-anomalous local circle, and determining a decision blockchain circle; through abnormal isolation control smart contract, node locking and decision switching are performed, and a decision instruction based on a multi-objective optimization algorithm is generated. The present application improves the abnormal self-healing ability and dynamic decision efficiency of the construction monitoring system.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] A TBM remote monitoring method based on a multi-level collaborative management architecture is applied to a remote integrated monitoring platform. The platform includes a multi-level collaborative management architecture comprising: a construction operation layer, a construction control layer, and a decision support layer. Each monitoring scenario's construction operation layer corresponds to multiple construction face nodes. Each monitoring scenario forms multiple local blockchain circles, each containing the corresponding multiple construction face nodes. These local blockchain circles form a unified blockchain circle, which corresponds to the decision support layer. Each decision support layer corresponds to a decision node. The remote monitoring method includes:

[0006] The construction status is monitored and construction status data is collected through the construction operation layer.

[0007] The construction status data is transmitted to the construction control layer, so that the construction control layer can perform construction status analysis and obtain analysis results.

[0008] Multiple analysis results are aggregated into the decision support layer, so that the decision support nodes in the decision support layer can determine the decision blockchain circle for the local blockchain circle whose current state is abnormal based on the multiple analysis results, wherein the decision blockchain circle is one of the local blockchain circles whose current state is not abnormal.

[0009] The construction control layer performs construction status analysis and obtains analysis results, including:

[0010] The construction status data is standardized to obtain equipment operating parameters, geological environment parameters, and construction efficiency parameters;

[0011] The equipment operating parameters are input into a preset equipment risk identification model, and the health status of the construction equipment and the risk status of the main bearing equipment are output through the equipment risk identification model. The equipment risk identification model includes a first model and a second model.

[0012] The geological environment parameters and construction efficiency parameters are input into a preset geological assessment model, and the surrounding rock grade and fracture zone distribution information are output through the geological assessment model.

[0013] The first model consists of a convolutional neural network and a long short-term memory neural network, which are used to output the health status of construction equipment. The input of the first model is the infrared temperature measurement image and vibration sequence signal in the equipment operating parameters.

[0014] The second model is based on a multilayer perceptron network, which is used to analyze the content and viscosity changes of ferromagnetic particles in lubricating grease to predict the risk status of the main bearing equipment. The input of the second model is the oil monitoring sensor data in the equipment operating parameters.

[0015] The core of the geological assessment model is a graph neural network, which includes a first branch and a second branch. The first branch performs a bandpass graph convolution operation, and the second branch performs a smooth graph convolution operation. The geological environment parameters and the construction efficiency parameters are input into the preset geological assessment model, including:

[0016] The construction surface nodes are mapped to monitoring points, wherein the node feature vectors of the monitoring points are calculated based on the geological environment parameters and the construction efficiency parameters;

[0017] Constructing an undirected graph structure based on the spatial location information of the monitoring points, wherein constructing the undirected graph structure includes: establishing a first-order edge for the monitoring point pair when the Euclidean distance between the monitoring points is less than a first distance threshold; and establishing a second-order edge for the monitoring point pair when the Euclidean distance between the monitoring points is greater than or equal to the first distance threshold and less than a second distance threshold.

[0018] The edge weights of the first-order and second-order edges are calculated based on the node feature vectors of the monitoring points to construct an undirected graph structure.

[0019] Based on the first branch, a bandpass graph convolution operation is performed on the first subgraph formed by the first-order edge to obtain the first feature matrix for lithological abrupt changes at the monitoring point.

[0020] Based on the second branch, a smooth graph convolution operation is performed on the second subgraph formed by the second-order edge to obtain a second feature matrix for the continuity of the monitoring point segment.

[0021] In a pre-defined multi-layer gating network, the first feature matrix and the second feature matrix are subjected to diffusion and aggregation operations in sequence according to the layer number;

[0022] After each layer of operation is completed, the degree of difference of the same monitoring point in the feature matrix is ​​calculated;

[0023] If the difference is greater than the preset difference threshold, then in subsequent layers, the element corresponding to the second feature matrix of the monitoring point will stop being updated, while the diffusion of the element corresponding to the first feature matrix will be retained.

[0024] If the difference is less than or equal to the difference threshold, then in subsequent layers, the element corresponding to the first feature matrix of the monitoring point will be stopped from being updated, and the aggregation of the element corresponding to the second feature matrix will be retained.

[0025] After all gating network operations are completed, select the corresponding elements in the feature matrix of each monitoring point that is still active to form a target feature matrix. Then, input the target feature matrix into the classification layer and output the surrounding rock grade result and the probability of fracture zone distribution of the monitoring point.

[0026] The diffusion and polymerization operations include:

[0027] When the layer number is odd, the first feature matrix is ​​locally diffused through the diffusion gating unit, and the updated first feature matrix of this layer is input to the second feature matrix through the attention mechanism as an additional input for the aggregation of the second feature matrix in the next layer. The second feature matrix is ​​the state of the matrix after the previous layer update.

[0028] When the layer number is even, the second feature matrix is ​​aggregated with segment information through the aggregation gating unit, and the updated second feature matrix of this layer is input to the first feature matrix through the attention mechanism as an additional input for the diffusion of the first feature matrix in the next layer. The first feature matrix is ​​the state of the matrix after the previous layer update.

[0029] The remote monitoring method further includes:

[0030] Summarize the consensus results transmitted within the aforementioned decision-making blockchain circle;

[0031] Perform a global state determination based on the consensus result;

[0032] Based on the global state, decision instructions are generated, wherein the generation of decision instructions is based on the NSGA-II algorithm, which is used to optimize objective conditions, including minimizing the escape time, minimizing resource scheduling costs, and maximizing the progress guarantee rate.

[0033] The determination of the decision blockchain circle for the local blockchain circle in the current state of anomaly includes:

[0034] Mark the abnormal construction surface node where the abnormal event occurred and its corresponding first local blockchain circle;

[0035] Based on the data flow relationship of all construction surface nodes maintained on the first partial blockchain circle, the propagation path of the first partial blockchain circle is deduced;

[0036] Calculate the reverse coverage capability of any local blockchain circle whose current state is not abnormal along the path of propagation in the first local blockchain circle.

[0037] The local blockchain circle with the largest reverse coverage capability is selected as the decision blockchain circle. If there are multiple local blockchain circles with the same reverse coverage capability, the local blockchain circle that is physically farthest from the first local blockchain circle is selected as the decision blockchain circle.

[0038] The remote monitoring method further includes:

[0039] Deploy anomaly isolation control smart contracts on-chain;

[0040] According to the path, the on-chain transaction permissions of the abnormal construction surface node and directly related nodes are initialized and locked through the abnormal isolation control smart contract;

[0041] Triggering each construction node within the decision blockchain circle to perform anomaly isolation consensus operation, so as to confirm the scope of anomaly control based on the consensus result;

[0042] After consensus is confirmed, the anomaly isolation control smart contract adjusts the control strategy for the abnormal construction surface node.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] 1. The remote monitoring method based on a multi-level collaborative management architecture provided by this invention adopts a layered design of construction operation layer, construction control layer and decision support layer. Combined with the data synchronization and decision collaboration mechanism between the local blockchain circle and the overall blockchain circle, it can effectively alleviate the data processing burden of the central server, realize the hierarchical processing and dynamic management of construction status data, and improve the flexibility and response speed of the remote integrated monitoring platform in complex construction environments. Attached Figure Description

[0045] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0046] Figure 1 This is a schematic diagram illustrating an exemplary application scenario of an embodiment of the present invention;

[0047] Figure 2 This is a flowchart illustrating a TBM remote monitoring method based on a multi-level collaborative management architecture, according to an embodiment of the present invention.

[0048] Figure 3 This is a flowchart illustrating an analysis result calculation method according to an embodiment of the present invention;

[0049] Figure 4 This is a network structure diagram of the geological assessment model in an embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram illustrating the process of the geological assessment model of this invention outputting information on the surrounding rock grade and fracture zone distribution.

[0051] Figure 6 This is a schematic diagram illustrating the diffusion and polymerization operation principles of an embodiment of the present invention;

[0052] Figure 7 This is a flowchart illustrating the decision blockchain circle determination method according to an embodiment of the present invention;

[0053] Figure 8 This is a flowchart illustrating the consensus method for decision-making in the blockchain community according to an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0055] Please see Figure 1 This figure is a schematic diagram of an exemplary application scenario provided by an embodiment of this application.

[0056] like Figure 1 As shown, the application scenario consists of a command center, a monitoring center, and a TBM, which correspond to the decision support layer, construction control layer, and construction operation layer, respectively.

[0057] Figure 1 The diagram shows that the TBM is located at the construction operation layer and is equipped with a variety of sensor devices to collect construction status information such as cutter head wear status, cutter head vibration, feed speed, surrounding rock pressure, and geological exploration data in real time, and performs preliminary standardization processing through a local processing module.

[0058] In one example, the sensor devices include, but are not limited to: a tool wear monitoring sensor for real-time detection of tool wear and cutter ring wear depth, supporting non-contact infrared sensing or contact displacement sensing technology; a cutterhead vibration sensor, employing an accelerometer or vibration velocity sensor, to monitor the amplitude, frequency, and direction changes of the cutterhead vibration during the tunneling process, assisting in the judgment of geological changes or mechanical failures; a thrust monitoring sensor, installed in the propulsion cylinder system, to measure the magnitude of the thrust applied by the propulsion cylinder to the TBM shield, providing real-time feedback on propulsion load fluctuations; a torque monitoring sensor, integrated into the main drive system, to measure changes in spindle torque to determine changes in cutterhead cutting resistance; a seepage pressure sensor, deployed around the shield to monitor changes in pore water pressure in the surrounding rock, used for early warning of geological disaster risks such as water inrush and mudslides; and a ground-penetrating radar detector for early detection. The system employs various sensors to monitor changes in the geological structure ahead of the tunnel face, such as faults, fracture zones, and karst caves, using GPR (Ground Ground Radar) technology. Temperature sensors are installed in key components such as the main bearings and hydraulic systems to monitor equipment operating temperatures in real time and prevent overheating failure. Slag discharge monitoring sensors, installed in the belt conveyor system, measure the amount of slag discharged per unit time to assist in assessing tunneling efficiency and tunnel face stability. Settlement monitoring sensors are used to detect the settlement and offset of the TBM body and tunnel lining structure in real time, improving the accuracy of tunneling attitude control. Spatial attitude sensors, such as gyroscopes and inclinometers, monitor changes in the TBM's tunneling direction and slope to ensure the accuracy of the tunnel axis. In practice, these sensor devices are connected to the local data processing module via wired or wireless data links and uploaded to the monitoring center for further analysis and data storage through standardized interfaces. The data from these sensors is not only used for routine operational monitoring but also serves as a crucial data source for subsequent construction status analysis, risk identification, and decision support.

[0059] Figure 1 The monitoring center corresponds to the construction control layer, receiving construction status data uploaded from multiple TBM cabs. Based on equipment operating parameters, geological environment parameters, and construction efficiency parameters, it calls preset equipment risk identification models and geological assessment models to complete a comprehensive analysis of equipment health status, main bearing equipment risk status, surrounding rock grade, and fracture zone distribution information, generating analysis results. Simultaneously, it maintains data notarization for each local blockchain circle on the blockchain. Figure 1 There are three monitoring centers, namely Monitoring Center 1, Monitoring Center 2 and Monitoring Center 3. In actual implementation, the number of monitoring centers can be determined according to the distribution range of the project.

[0060] Figure 1The diagram illustrates the decision support layer corresponding to the command center. It aggregates analysis results transmitted from multiple local blockchain circles and executes intelligent decisions through decision support nodes, including anomaly detection, selection of decision blockchain circles, generation and issuance of decision instructions. Based on the consensus mechanism formed by the overall blockchain circle, the command center can coordinate with various operation management centers and the TBM control room to achieve rapid response and optimized resource scheduling in abnormal situations.

[0061] Next, with reference to the accompanying drawings, we will introduce a TBM remote monitoring method based on a multi-level collaborative management architecture provided by an embodiment of this application.

[0062] Please see Figure 2 This figure is a flowchart illustrating a TBM remote monitoring method based on a multi-level collaborative management architecture provided in this application embodiment. It is applied to a remote integrated monitoring platform. The platform includes a multi-level collaborative management architecture, comprising: a construction operation layer, a construction control layer, and a decision support layer. Each monitoring scenario's construction operation layer corresponds to multiple construction face nodes. Each monitoring scenario forms multiple local blockchain circles, each containing multiple corresponding construction face nodes. These local blockchain circles form a unified blockchain circle, which corresponds to the decision support layer. The decision support layer corresponds to a decision node. The remote monitoring method includes steps S1-S3, with the specific steps as follows:

[0063] S1: Monitor construction status through the construction operation layer and collect construction status data;

[0064] In this embodiment, the construction operation layer consists of multiple construction face nodes. Each construction face node is equipped with several types of sensor devices to collect real-time operating status parameters of the operating equipment, operating environment status parameters, and operation efficiency parameters. These sensor devices include, but are not limited to, wear status sensors, vibration status sensors, thrust measurement sensors, environmental pressure sensors, and geological change detectors. The collected construction status data undergoes preliminary standardization processing through a local processing module, including outlier removal, data normalization, and unified feature encoding. The data is then stored on the corresponding local blockchain in the form of data transaction evidence. This achieves efficient collection, standardized processing, and on-chain storage of construction site status information, ensuring the real-time nature, integrity, and immutability of the data source.

[0065] S2: Transmit the construction status data to the construction control layer so that the construction control layer can perform construction status analysis and obtain the analysis results;

[0066] In this embodiment, the construction control layer receives data transactions from multiple construction face nodes and further extracts key indicators of the construction status based on a preset data processing strategy. The construction control layer invokes multiple preset intelligent analysis models. One model, based on a fusion structure of convolutional neural networks and long short-term memory neural networks, takes operating parameters of the working equipment as input and outputs the health status of the equipment and the risk status of auxiliary equipment. Another model, based on a multilayer perceptron network, analyzes parameters collected by oil monitoring sensors to infer the health status of the lubrication system. Simultaneously, a geological assessment model, based on a graph neural network structure, combines operating environment parameters and operating efficiency parameters to output information on the surrounding rock type and potential fracture zone distribution at the working face. Through the collaborative operation of multiple models, refined analysis and intelligent risk identification of key states during the construction process can be achieved, improving the construction control layer's responsiveness and predictive capabilities to changes in operating status.

[0067] S3: Summarize multiple analysis results to the decision support layer, so that the decision support nodes in the decision support layer can determine the decision blockchain circle for the local blockchain circle with the current abnormal state based on multiple analysis results;

[0068] In this embodiment, the decision support layer receives analysis results uploaded from each local blockchain circle. When an abnormal construction node is detected in a local blockchain circle, the decision support node deduces the anomaly propagation path based on the node data flow relationship recorded on the chain, and calculates the reverse coverage capability of each of the non-abnormal local blockchain circles for the anomaly propagation path. The local blockchain circle with the largest reverse coverage capability is selected as the decision blockchain circle first. If there are multiple local circles with the same reverse coverage capability, the local circle with the largest distance from the abnormal circle is further selected based on the physical location information. Through the reverse anomaly isolation mechanism, the decision blockchain circle with the strongest anomaly suppression capability can be quickly screened out, ensuring that in the event of a sudden anomaly, the remote integrated monitoring platform can stably and efficiently complete decision takeover and status control, thereby improving the emergency response capability and on-chain collaborative control level of the overall construction management system.

[0069] While existing technologies utilize distributed systems for remote monitoring data synchronization and anomaly detection, they still suffer from fundamental shortcomings in complex and dynamic construction environments. First, current technologies often employ a single-level, flat data processing model, where all monitoring nodes synchronize data and make collaborative decisions at the same level. In highly heterogeneous and anomaly-prone construction environments, this architecture easily leads to complex decision-making chains, increased system load, and significantly increased response latency, failing to meet the demands for rapid, accurate, and hierarchical anomaly handling. Second, traditional anomaly decision-making models typically rely on the health status of a single node or fixed indicator scoring for screening, lacking dynamic perception and reverse control of anomaly propagation paths. This fails to effectively block the anomaly spread chain, causing local anomalies to amplify their impact on the entire system, reducing the reliability and stability of the remote monitoring platform. Furthermore, even when blockchain evidence storage mechanisms are introduced, existing solutions are limited to data tamper-proofing and fail to fully utilize the blockchain's circular structure for intelligent collaborative decision-making and autonomous response in anomaly situations. The overall design exhibits significant systemic flaws and a lack of innovation.

[0070] This application proposes a TBM remote monitoring method based on a multi-level collaborative management architecture, forming a systematic innovation from architecture to decision-making mechanism. In the remote integrated monitoring platform, this application constructs a multi-level collaborative management architecture consisting of a construction operation layer, a construction control layer, and a decision support layer. Each layer is responsible for data collection, intelligent analysis, and decision command generation, respectively. Data synchronization and consensus collaboration between layers are achieved through a blockchain circle. Each node in the construction operation layer constitutes a local blockchain circle, and multiple local circles are merged to form the overall blockchain circle. The decision support layer centrally controls anomaly detection and decision migration through the overall blockchain circle. This architecture not only achieves hierarchical management and optimization of data flow, control flow, and resource flow, but also ensures the credibility and traceability of data flow and decision-making processes between layers through blockchain technology. Compared to traditional single-level centralized decision-making, the multi-level collaborative architecture of this application can quickly decentralize decision-making power to healthy local blockchain circles in the event of local anomalies, greatly reducing the processing complexity of the central node, improving the flexibility of decision-making and system resilience, and solving major problems such as excessively long decision chains and processing delays caused by the single architecture in existing technologies.

[0071] Furthermore, this application introduces a dynamic decision circle switching mechanism based on anomaly propagation path derivation and reverse coverage capability screening during the anomaly detection and decision circle determination process. Specifically, after detecting a local circle anomaly, the anomaly propagation path is deduced based on the data flow relationship of construction surface nodes maintained on the chain, the reverse coverage capability of other healthy circles on the anomaly link is evaluated, and the local circle that can maximally block the spread of anomalies is preferentially selected as the decision circle. Simultaneously, during the feature extraction process of construction status data, this application proposes a gated multi-layer processing mechanism with alternating diffusion and aggregation of odd and even layers, dynamically adjusting the weights of local mutation features and segment continuity features. Through difference-driven feature freezing technology, the problems of excessive smoothing or noise amplification in traditional graph convolution models are avoided.

[0072] Please see Figure 3 The figure is a flowchart illustrating an analysis result calculation method provided in an embodiment of this application. Figure 3 The method shown can be applied to step S2 of the aforementioned method, and the specific steps are as follows:

[0073] S2.1: Standardize the construction status data to obtain equipment operating parameters, geological environment parameters, and construction efficiency parameters;

[0074] Specifically, due to the diverse sources, inconsistent sampling frequencies, high noise levels, and different units of measurement in the raw data collected from construction sites, directly using this data for subsequent analysis would lead to model input distortion and affect the accuracy of the analysis. Therefore, this step requires standardization of the raw construction status data. Standardization includes, but is not limited to, the following operations: normalizing continuous data to a specified interval (e.g., between 0 and 1), performing one-hot encoding on discrete data, using sliding window interpolation or nearest neighbor interpolation to complete missing data, and using median substitution or local weighted regression to correct outliers. Simultaneously, the units of measurement are standardized, for example, pressure units are standardized to kPa, temperature units to degrees Celsius, and length units to meters, to ensure consistency across different data sources. Through this standardization process, the data input to each model has consistent feature dimensions, comparable scales, and noise controlled within acceptable limits, enhancing the stability and accuracy of model training and inference, and improving the reliability of construction status analysis.

[0075] S2.2: Input the equipment operating parameters into the preset equipment risk identification model, and output the health status of the construction equipment and the risk status of the main bearing equipment through the equipment risk identification model, wherein the equipment risk identification model includes a first model and a second model;

[0076] Specifically, due to the complex operating status of construction equipment and its significant influence from environmental factors, a single parameter cannot comprehensively assess the equipment's health status. Therefore, in this step, standardized equipment operating parameters are input into two risk identification models with different structures to achieve multi-angle, multi-scale risk assessment. The first model is primarily a concatenated structure of a Convolutional Neural Network (CNN) and a Long Short-Term Memory Neural Network (LSTM). The CNN part is mainly responsible for extracting local heating patterns and hotspot distribution features from infrared thermography images, while the LSTM part performs time-series modeling of vibration sequence signals to capture periodic changes, short-term impacts, or persistent abnormal trends that occur during equipment operation. By jointly extracting spatial and temporal dynamic features, the first model can comprehensively judge the overall health status of the construction equipment, such as whether there are potential overheating faults, bearing wear, or structural fatigue problems.

[0077] In one example, the first model of this embodiment includes:

[0078] The infrared temperature measurement image and vibration sequence signal in the operating parameters of the receiving device are used as input data. The infrared temperature measurement image is normalized to a standard format of 224×224 pixels, and the vibration sequence signal is extracted in a 2-second time window and sampled at a fixed sampling frequency of 500Hz. This unifies the dimension and scale of the input data and ensures the consistency and effectiveness of subsequent network processing.

[0079] Furthermore, firstly, for the input infrared temperature measurement image, the first model includes a convolutional neural network module. The CNN module consists of four sets of convolutional layers, batch normalization layers, activation layers, and pooling layers stacked together. Each set of convolutional layers uses 32, 64, 128, and 256 convolutional kernels, respectively, with a kernel size of 3×3 and a stride of 1. The boundaries are padded in the same way to keep the feature map size constant. Each convolutional layer is followed by a batch normalization layer to accelerate training convergence and improve the model's generalization ability, followed by a ReLU activation function to introduce non-linear feature mapping. A max pooling layer is set after every two sets of convolutions, with a pooling window size of 2×2 and a stride of 2, gradually reducing the feature map size and extracting important local features from the image. Through the above multi-layer convolution and pooling operations, the local heating pattern, hotspot distribution characteristics, and overall temperature rise trend in the infrared image are extracted to assist in the subsequent determination of the surface thermal anomaly status of the construction equipment.

[0080] Furthermore, for vibration sequence signal input, the first model includes a combined module of a one-dimensional convolutional network and a long short-term memory (LSTM) neural network. First, local time-domain features are extracted using a two-layer one-dimensional convolutional network. The first layer uses 32 convolutional kernels with a kernel size of 5 and a stride of 1; the second layer uses 64 convolutional kernels with a kernel size of 3 and a stride of 1. After convolution, batch normalization layers and ReLU activation functions are connected to accelerate feature convergence and enhance nonlinear modeling capabilities. The convolutional output is then input to a stacked two-layer LSTM module, with 128 hidden units per layer. This bidirectional LSTM structure simultaneously captures the forward and backward time-dependent features of the vibration signal, enhancing the ability to identify periodic impacts and sudden abnormal vibrations. The LSTM module can long-term memorize trend changes and short-term abnormal signals in the equipment vibration signal, improving the sensitivity and accuracy of vibration sequence anomaly detection.

[0081] Furthermore, after feature extraction is completed in their respective modules, the infrared thermography image processing path and the vibration sequence signal processing path perform feature concatenation in the feature fusion layer. The feature fusion layer uses a simple join operation to concatenate the two feature vectors along their respective feature dimensions into a unified feature vector. After fusing the feature vector, it is input to a fully connected layer containing 512 nodes. Following the fully connected layer are batch normalization layers, ReLU activation functions, and Dropout layers for further feature compression, nonlinear transformation, and overfitting prevention.

[0082] Finally, the feature vectors are fed into the output layer, which is a classification layer with a softmax activation function. It outputs probability distributions for three categories, corresponding to the normal, slightly abnormal, and severely abnormal states of the construction equipment, respectively. It also outputs a continuous confidence score for subsequent threshold judgment and policy adjustment. The output layer uses a cross-entropy loss function during training and the Adam optimizer for parameter updates. The initial learning rate is set to 0.001, and the learning rate is dynamically adjusted based on the validation set performance during training.

[0083] Specifically, since the condition of the lubrication system of construction equipment is affected by a variety of minute factors, and the deterioration process often manifests as subtle changes in multiple indicators in the early stages, relying solely on a single oil monitoring parameter cannot comprehensively assess the health status of the lubrication system. Therefore, in this step, the oil monitoring sensor data from the standardized equipment operating parameters is input into the second model for analysis to achieve a multi-faceted and comprehensive assessment of the lubrication system's condition. The second model is primarily a multilayer perceptron (MLP) network structure. Through deep nonlinear feature mapping, it can automatically extract key physicochemical features such as ferromagnetic particle content change curves and viscosity change curves from the original oil monitoring data corresponding to the lubricating grease. Oil monitoring data includes, but is not limited to, temperature, conductivity, particle concentration, viscosity measurements, and liquid contaminant concentration. Through feature extraction and combination using the MLP multilayer structure, it can effectively capture subtle correlations between different oil properties, thereby inferring the potential deterioration trend of the lubrication system. By constructing a multilayer feature interaction mapping space, the second model can comprehensively identify early risk issues such as lubricating oil contamination, decreased lubrication performance, and system leakage, thereby outputting the corresponding risk status of the main bearing equipment. Compared with the traditional method that relies solely on static thresholds for a single oil parameter, the second model using an MLP network can capture abnormal signals in the early stages of equipment operation, improving the sensitivity of anomaly detection and the timeliness of early warning, and effectively supporting the intelligent early identification and dynamic response of lubrication system faults by the remote integrated monitoring platform.

[0084] In one example, the second model of this embodiment includes:

[0085] First, an input layer is set up to receive the standardized raw data from the oil monitoring sensor. The input dimension is determined based on the number of collected parameters; for example, if there are 5 key parameters, the input vector length is 5. The input layer is followed by a first hidden layer containing 512 nodes. This first hidden layer uses a fully connected approach, with each node followed by a ReLU activation function. This is used to achieve a preliminary linear combination and nonlinear mapping of the raw oil features, capturing the changing relationships between basic features.

[0086] The output of the first hidden layer is connected to a second hidden layer, which contains 256 nodes. This second hidden layer continues with a fully connected approach and a ReLU activation function to further explore higher-order feature interaction information, especially the deep correlation between the potential change patterns of ferromagnetic particle content and the implicit change trend of viscosity. The role of the second hidden layer is to implicitly map the original oil state parameters to the feature space of the ferromagnetic particle content change curve and the viscosity change curve.

[0087] After the output of the second hidden layer, two parallel third-layer output submodules are connected:

[0088] The first submodule outputs the characteristics of ferromagnetic particle content changes, and the second submodule outputs the characteristics of viscosity changes. Each submodule consists of a set of 64 fully connected layers and a linear activation function, used to regress and infer the growth trend of ferromagnetic particles and the viscosity evolution trend in the time series, respectively. Through this parallel feature inference structure, the changes of the two most critical indicators in the process of lubricating oil deterioration can be extracted independently and accurately.

[0089] The feature results output from the two submodules are then jointly analyzed in the fusion layer, which consists of a 128-node fully connected layer using the ReLU activation function. The fused feature vectors then enter the final risk output layer, a single-node structure using the Sigmoid activation function. The output value is a continuous probability value between 0 and 1, representing the potential risk level of the main bearing equipment. When the risk output value exceeds a preset threshold, the lubrication system anomaly warning logic is automatically triggered, assisting the remote integrated monitoring platform in responding promptly.

[0090] Furthermore, during the model training phase, the weighted mean squared error (MSE) is used as the loss function. Weights are assigned to the prediction errors of ferromagnetic particle changes, viscosity changes, and the final risk output error. These weights are dynamically adjusted based on the frequency of anomalous samples in the training set to enhance the detection capability of rare anomalous patterns. The optimizer uses Adam with an initial learning rate of 0.0005, and an Early Stopping mechanism is introduced to terminate training early if there is no improvement in performance on the validation set, preventing overfitting.

[0091] S2.3: Input the geological environment parameters and the construction efficiency parameters into the preset geological assessment model, and output the surrounding rock grade and fracture zone distribution information through the geological assessment model;

[0092] Specifically, the stability of the surrounding rock and the distribution of fracture zones in the construction environment are key factors affecting construction safety and efficiency. Traditional methods rely on manual experience for judgment, which suffers from strong subjectivity and delayed response. Therefore, in this step, standardized geological environment parameters (such as surrounding rock pressure, seepage pressure, and geological advance detection signal characteristics) and construction efficiency parameters (such as advance speed, unit energy consumption, and slag discharge) are used as inputs and imported into a geological assessment model for intelligent inference. The geological assessment model is based on a graph neural network (GNN) structure. The model uses construction face nodes as vertices, and the node feature vectors are generated by encoding the aforementioned input parameters. The nodes are connected to each other through Euclidean distance and construction segment relationships to form an undirected graph. First-order edges and second-order edges are established to model local geological abrupt changes and large-scale segment continuity relationships, respectively.

[0093] Taking the network structure of a geological assessment model as an example, you can refer to... Figure 4 To understand, Figure 4The geological assessment model network structure diagram provided in this application embodiment is shown. The model performs bandpass graph convolution on the first-order edge subgraph in the first branch to highlight local lithological change signals. In the second branch, it performs smoothing graph convolution on the second-order edge subgraph to enhance the overall continuity characteristics of the surrounding rock. The features of the two branches are processed alternately according to layer number in a multi-layer gated network. When the layer number is odd, a diffusion operation is performed based on local abrupt changes, and the information is injected into the overall continuity feature matrix. When the layer number is even, an aggregation operation is performed based on the overall continuity features, and the information is injected into the local feature matrix, alternately promoting the dynamic fusion and supplementation of local anomaly information and overall background information. After each layer, the model calculates the difference index of the same monitoring point in the two feature paths and dynamically freezes the feature flow direction based on the difference: if the difference is higher than a preset threshold, the overall continuity branch features are stopped from being updated, and only the local abrupt change feature flow is retained; if the difference is lower than or equal to the threshold, the opposite occurs: local branch updates are stopped, and the overall feature flow is retained. Finally, after all gating operations are completed, the model selects active feature matrix elements based on the frozen state to form the target feature matrix. The target feature matrix is ​​input into the classification layer, which generates the surrounding rock grade label and the probability distribution of fracture zone occurrence for each monitoring point through a multi-output mechanism, thereby achieving high-precision intelligent prediction of the geological state of the monitoring point under complex geological change environment.

[0094] In one example, since the input to the geological assessment model in this embodiment is a graph structure, the geological environment parameters and construction efficiency parameters need to be preprocessed to generate an undirected graph structure. The specific processing procedure is as follows:

[0095] The construction surface nodes are mapped to monitoring points, wherein the node feature vectors of the monitoring points are calculated based on the geological environment parameters and the construction efficiency parameters;

[0096] Specifically, since the original construction status data is collected in a distributed manner with construction face nodes as the basic unit, and the geological environment parameters and construction efficiency parameters collected from different construction face nodes are spatially different, it is necessary to map each construction face node to a unique monitoring point for subsequent unified processing.

[0097] In this embodiment, the node feature vector of each monitoring point is obtained by feature encoding of geological environmental parameters (such as surrounding rock strength, rock mass integrity index, porosity, and permeability coefficient) and construction efficiency parameters (such as advance speed, unit energy consumption, and construction time interval) collected from the corresponding construction face node. Specifically, continuous parameters are processed using standard normalization, while discrete parameters are represented by one-hot encoding. All parameters are concatenated into a fixed-dimensional feature vector. This ensures that the features of different monitoring points are expressed in the same feature space, facilitating subsequent node relationship modeling based on feature similarity. It also reduces the interference of feature deviations caused by sampling differences on the graph structure learning results, improving the model's adaptability to complex geological changes.

[0098] Constructing an undirected graph structure based on the spatial location information of the monitoring points, wherein constructing the undirected graph structure includes: establishing a first-order edge for the monitoring point pair when the Euclidean distance between the monitoring points is less than a first distance threshold; and establishing a second-order edge for the monitoring point pair when the Euclidean distance between the monitoring points is greater than or equal to the first distance threshold and less than a second distance threshold.

[0099] In this embodiment, since the spatial relationship between different monitoring points directly reflects geological continuity or local abrupt changes, edge connections between monitoring points are established using spatial location information. To this end, the Euclidean distance between any two points is first calculated based on the three-dimensional coordinates of each monitoring point. Two threshold parameters are set for the calculated distance values: a first distance threshold D1 and a second distance threshold D2, where D1 is less than D2. If the Euclidean distance between two monitoring points is less than D1, a strong local correlation is considered between the two points, and a first-order edge is established; if the Euclidean distance is between D1 and D2, a weak correlation is considered between the two points, and a second-order edge is established. The first-order edge is used to capture local abrupt changes, such as rapid changes in the surrounding rock grade within a small area, while the second-order edge is used to maintain the continuity of the overall geological environment within the region. This hierarchical edge establishment strategy allows the undirected graph structure to efficiently model local change information and capture segment stability information, avoiding the problems of excessive smoothing or loss of detail caused by a single adjacency scale.

[0100] The edge weights of the first-order and second-order edges are calculated based on the node feature vectors of the monitoring points, and an undirected graph structure is constructed.

[0101] Specifically, for each pair of connected monitoring points that have been established, the cosine similarity value between their node feature vectors is calculated, and the normalized cosine similarity is used as the initial weight of the edge.

[0102] Please see Figure 5 The figure is a schematic diagram of the process by which the geological assessment model provided in this application outputs information on the surrounding rock grade and fracture zone distribution. Figure 5The method shown can be applied to step S2.3 of the aforementioned method, and the specific steps are as follows:

[0103] S2.3.1: Perform a bandpass graph convolution operation on the first subgraph formed based on the first-order edge according to the first branch to obtain the first feature matrix for lithological abrupt changes at the monitoring point;

[0104] Specifically, in complex construction environments, local lithological abrupt changes often manifest as high-frequency, small-scale geological feature variations between monitoring points. These changes have a significant impact on construction safety and progress. Traditional graph convolution operations based on mean aggregation, due to excessive smoothing, easily lead to the neglect of these subtle but important features. To highlight local lithological variations, this step employs a bandpass graph convolution operation for subgraphs formed based on first-order edges.

[0105] In this embodiment, local neighborhoods are first extracted from the subgraph established by the first-order edges. The adjacency matrix is ​​processed through local normalization to standardize the node degree information. In the design of the bandpass convolution kernel, bandpass enhancement is performed to address the feature differences between a node and its neighboring nodes, extracting the changes in node features in the mid-frequency band. Specifically, this is achieved by performing a higher-order Laplacian operator transformation on the node features, filtering feature change components whose rate of change falls within a specific bandwidth (for example, a 15%-40% change range). During convolution, specific coefficients are used to strengthen features within the set change range while weakening the propagation of features from other regions. These specific coefficients can be determined experimentally by those skilled in the art. By highlighting local lithological abrupt change signals in the feature space while suppressing the interference of the overall low-frequency change trend on local detail detection, the resulting first feature matrix contains rich local feature information on lithological abrupt changes.

[0106] S2.3.2: Perform a smooth graph convolution operation on the second subgraph formed based on the second-order edge according to the second branch to obtain the second feature matrix for the continuity of the monitoring point segment;

[0107] Specifically, in order to capture the continuous changing trend of the surrounding rock over a large area within the construction section, a smooth graph convolution operation needs to be performed on the subgraph established by the second-order edge. The essential characteristic of section continuity is that the features of neighboring nodes change little and the node states are similar. Therefore, in this embodiment, a second-order edge subgraph is established for node pairs whose Euclidean distance is between the first threshold and the second threshold.

[0108] In this embodiment, a low-pass filter convolution kernel design is adopted during the smooth convolution process. This involves introducing adjacency matrix normalization and feature consistency enhancement mechanisms based on standard convolution operations. Specifically, during convolution calculation, a weighted average is performed on the node features and the features of neighboring nodes. The weighting coefficients are based on node feature similarity (e.g., cosine similarity), with higher similarity nodes receiving greater weights and lower similarity nodes receiving less weight, dynamically adjusting information aggregation. This enhances the expression of the coherent geological structure of the section, and the resulting second feature matrix primarily characterizes the continuously changing geological environment state of the monitoring point along the tunnel axis or construction route.

[0109] S2.3.3: In the preset multi-layer gated network, the first feature matrix and the second feature matrix are subjected to diffusion and aggregation operations in sequence according to the layer number;

[0110] Specifically, in the geological environment of construction, local abrupt changes (such as fracture zones and small faults) and overall continuity (such as the transition between soft and hard surrounding rock) often coexist and are interspersed. Modeling solely based on either local or overall features will lead to the loss of feature information or a decrease in anomaly detection accuracy. If only local feature diffusion is considered, the overall geological trend is easily ignored, resulting in insufficient understanding of the continuity of long-distance tunnel excavation; if only overall feature aggregation is performed, important local abrupt changes are easily smoothed out. Therefore, this embodiment proposes a mechanism in a multi-layer gated network to alternately perform diffusion and aggregation operations based on odd and even layers, in order to achieve a dynamic balance between local sensitivity and overall continuity.

[0111] In this embodiment, when the current convolutional layer number is odd, the diffusion operation based on the first feature matrix (local mutation feature matrix) is executed first. The essence of the diffusion operation is to utilize the local change features of nodes and propagate them along the first-order edges of the adjacency matrix within the neighborhood. However, an attention weighting mechanism is introduced during the diffusion process, allowing local mutation features to selectively influence surrounding nodes. Specifically, the attention mechanism dynamically assigns weights based on the feature similarity between the source and target nodes, ensuring that only node pairs with geological anomaly correlations are amplified in the diffusion, rather than spreading indiscriminately. This aims to simulate the phenomenon that geological anomalies typically extend along weak geological surfaces such as fracture zones during actual tunnel excavation, thus enabling local mutation features to diffuse reasonably and locally, rather than spreading disorderly to unrelated areas.

[0112] Furthermore, at even-numbered layers, an aggregation operation based on the second feature matrix (overall continuity feature matrix) is performed. This aggregation operation uses the adjacency matrix to weighted average the features of nodes within the neighborhood, strengthening the continuity information of the segments. To ensure that local anomalies are not completely masked, this embodiment introduces local mutation information output from the first feature matrix of the previous layer as an auxiliary input during the aggregation process, adjusting the aggregation weight of each node. If a node has a high mutation signal in its local features, it is given a smaller weight during the overall feature aggregation process to avoid excessive dilution of local anomalies by the overall trend. Conversely, if the degree of local mutation in a node is small, the aggregation weight participates normally, allowing the continuous features to be fully preserved.

[0113] Through this alternating diffusion and aggregation mechanism of odd and even layers, a dynamic and adaptive information flow process is essentially constructed. In areas with significant local mutations, the system automatically strengthens the propagation of local information, forming a localized enhancement of features centered on anomalies; while in areas with strong geological continuity, the system tends to aggregate overall features consistently, maintaining a smooth expression of the overall geological structure. This processing logic conforms to the natural law of localized outbreaks of geological anomalies and overall stability in normal areas during construction sites, enabling the model to adaptively adjust its feature extraction strategy under different geological structures.

[0114] Preferably, to enhance the rationality and stability of this alternating diffusion-aggregation process, this embodiment introduces a dynamic mask module into the gating mechanism. The dynamic mask adjusts the switching of diffusion and aggregation paths in real time based on the intensity of local mutations at nodes, achieving node-level feature selection and ensuring that the local-to-global alternation process is not a rigid switch, but rather a dynamic and flexible adjustment based on data features. Overall, through the above design, the method of this application can maintain a high sensitivity to local anomalies under complex geological conditions, while maintaining a consistent expression of the geological characteristics of the entire tunneling section, significantly improving the comprehensive accuracy and reliability of surrounding rock grade prediction and fracture zone identification.

[0115] For example, regarding diffusion and polymerization operations, please refer to... Figure 6 To understand, Figure 6 This is a schematic diagram illustrating the diffusion and aggregation operation principle provided in this application embodiment. When the layer number is odd, the first feature matrix is ​​locally diffused through the diffusion gating unit, and the updated first feature matrix of this layer is input to the second feature matrix through an attention mechanism as an additional input for the aggregation of the second feature matrix in the next layer. The second feature matrix is ​​the state of the matrix after the update in the previous layer. When the layer number is even, the second feature matrix is ​​segmented through the aggregation gating unit, and the updated second feature matrix of this layer is input to the first feature matrix through an attention mechanism as an additional input for the diffusion of the first feature matrix in the next layer. The first feature matrix is ​​the state of the matrix after the update in the previous layer.

[0116] S2.3.4: After each layer of operation is completed, calculate the degree of difference of the same monitoring point in the feature matrix;

[0117] Specifically, since each monitoring point retains independent feature representations in both the first and second feature matrices after alternating diffusion and aggregation operations, it is necessary to measure the degree of feature difference for the same node in the two feature matrices in real time in order to determine whether the current node should be classified as a local anomaly or an overall trend. Therefore, in this step, the degree of difference is calculated for each monitoring point after each layer of diffusion or aggregation operation.

[0118] In this embodiment, Euclidean distance is used as the basic index. Specifically, the Euclidean distance is calculated between the eigenvectors of the first and second feature matrices for the same node, and the result is normalized to a fixed interval. Normalization can employ max-min value normalization or distribution standardization to ensure the comparability and stability of the difference index across feature matrices of different scales. A larger difference index indicates a more significant difference between local abrupt changes and overall continuous features, highlighting local anomalies at the node; a smaller difference index indicates that overall geological continuity dominates at the node. Introducing the difference index provides data support for subsequent node feature freezing strategies, enabling dynamic judgment based on actual feature changes rather than pre-set static rules, thus improving the model's adaptability and resilience to complex geological changes.

[0119] S2.3.5: If the difference is greater than the preset difference threshold, then in subsequent layers, the element corresponding to the second feature matrix of the monitoring point will be stopped from being updated, while the diffusion of the element corresponding to the first feature matrix will be retained.

[0120] Specifically, when the difference exceeds a set threshold, it indicates that the local anomaly characteristics of the monitoring point are dominant. If smoothing aggregation is continued on the second feature matrix, important local mutation features may be smoothed out by the overall trend, reducing the accuracy of local anomaly detection. Therefore, in this embodiment, for such nodes, their updates on the second feature matrix path are actively frozen in subsequent convolution operations, retaining only the continued diffusion on the first feature matrix path.

[0121] In this embodiment, the freezing is implemented by introducing a node-level mask matrix. Before the convolution operation, a freeze mask is generated based on the degree of difference. The convolution input of the marked node on the smooth path is set to zero or passed with an identity, thereby skipping the weight update of that path. In this way, it can be ensured that local anomalous signals are effectively preserved and enhanced in the multi-layer convolution process, without being diluted or lost in the subsequent overall feature aggregation process.

[0122] S2.3.6: If the difference is less than or equal to the difference threshold, then in subsequent layers, the element corresponding to the first feature matrix of the monitoring point is stopped from being updated, and the aggregation of the element corresponding to the second feature matrix is ​​retained;

[0123] Specifically, when the difference is lower than or equal to the threshold, it indicates that the local features of the node differ little from the overall trend, meaning the node is located in a region dominated by geological continuity. Continuing local diffusion will not significantly improve feature expression and may even introduce local noise. Therefore, in this embodiment, for such nodes, the updates on the first feature matrix path are actively frozen during subsequent convolution processes, retaining only the aggregated updates on the second feature matrix path.

[0124] In this embodiment, the freezing operation is also based on node mask matrix control, performing masking processing on the convolutional input of locally abrupt paths. This effectively focuses model resources on optimizing overall trend features, strengthens the expression of segment coherence features, and reduces the risk of local overfitting.

[0125] S2.3.7: After all gated network operations are completed, select the corresponding elements in the feature matrix of each monitoring point that is still active to form a target feature matrix, and input the target feature matrix into the classification layer to output the surrounding rock grade result and the probability of fracture zone distribution of the monitoring point.

[0126] Specifically, after all gating network layer operations are completed, for each monitoring point, based on the frozen records, the final feature vector of the point that is still active (unfrozen path) is selected as the output to form the target feature matrix. That is, if a node is active in the first feature matrix path, its first matrix feature is selected; if it is active in the second feature matrix path, its second matrix feature is selected.

[0127] Furthermore, the target feature matrix is ​​then input into the classification layer, which contains two output branches: one for rock grade classification, using a fully connected layer connected to a Softmax multi-class activation function to predict and classify the rock grade; and the other for fracture zone distribution prediction, using a fully connected layer connected to a Sigmoid activation function to output continuous values ​​of the probability of fracture zone occurrence. During training, the classification layer employs a joint loss function: cross-entropy loss for rock grade and binary cross-entropy loss for fracture zone distribution, with weights allocated according to task importance for joint optimization.

[0128] By dynamically freezing active path features and classifying them independently, this embodiment can adaptively output information on surrounding rock and fracture zones based on the dominant direction of actual node features. This enables efficient perception and accurate prediction of both surrounding rock stability and local fracture risk in complex construction geological environments, greatly enhancing the monitoring platform's real-time perception and response capabilities to construction safety risks.

[0129] Please see Figure 7 The figure is a flowchart illustrating the decision blockchain circle determination method provided in an embodiment of this application. Figure 7 The method shown can be applied to step S3 of the aforementioned method, and the specific steps are as follows:

[0130] S3.1: Mark the abnormal construction surface node where the abnormal event occurred and its first local blockchain circle;

[0131] Specifically, in the multi-level collaborative remote monitoring architecture, construction status data is synchronized on the blockchain by distributed construction face nodes, and each local blockchain circle manages the construction monitoring tasks within its local area. When an abnormal event occurs (such as equipment malfunction or geological anomaly), if the source of the anomaly cannot be identified and isolated in a timely manner, it can easily lead to the spread of the anomaly, affecting the decision-making efficiency and accuracy of the overall monitoring platform. Therefore, in this step, it is necessary to first mark the abnormal construction face node corresponding to the abnormal event on the blockchain and determine its first local blockchain circle.

[0132] In this embodiment, the marking of abnormal construction face nodes is based on the analysis results of the equipment risk identification model and the geological assessment model. When the abnormal state of a node exceeds a set threshold, an abnormal label is automatically triggered and written to the blockchain. Subsequently, by querying the on-chain construction face node-local blockchain circle mapping table, the local blockchain circle to which the abnormal node belongs is located as the first local blockchain circle. By marking the relationship between abnormal nodes and circle affiliation on the chain, a clear and accurate data foundation can be provided for subsequent deduction of abnormality propagation paths and decision circle selection, ensuring the timeliness and reliability of abnormality source identification.

[0133] S3.2: Based on the data flow relationship of all construction surface nodes maintained on the first partial blockchain circle, deduce the propagation path of the first partial blockchain circle;

[0134] Specifically, in construction monitoring operations, there are natural data interactions and operational linkages between various construction site nodes, such as sensor data streams and equipment control command streams. Abnormal states often propagate along the data flow path, causing adjacent nodes to be affected synchronously. Therefore, in this step, it is necessary to deduce the possible propagation paths of anomalies based on the data flow relationships of construction site nodes recorded on the first local blockchain.

[0135] In this embodiment, the data flow relationship is maintained through the on-chain transaction call relationship between construction surface nodes. When there is direct transaction interaction or data synchronization behavior between nodes, directed edge information is recorded. When deriving the propagation path, the path recursively expands along the directed edges starting from the abnormal construction surface node until it can no longer expand or reaches the propagation depth limit, forming a set of abnormal propagation paths. Each propagation path includes not only the node sequence but also characteristic indicators such as transaction interaction frequency and transaction content type (e.g., equipment control, monitoring synchronization). By deriving the abnormal propagation link based on actual data flow, rather than simply speculating based on physical location or adjacency, the true diffusion trajectory of the abnormal state can be accurately depicted, avoiding over-isolation or under-isolation, and improving the targeting and accuracy of abnormal control.

[0136] S3.3: Based on any local blockchain circle whose current state is not abnormal, calculate its reverse coverage capability on the path of propagation in the first local blockchain circle;

[0137] Specifically, after an anomaly occurs, in order to quickly select a new decision-making circle within the healthy local blockchain circles that can effectively cut off the anomaly propagation path and assume decision-making responsibility, it is necessary to assess the reverse coverage capability of all healthy local circles. Reverse coverage capability reflects the controllability and intervention potential of nodes within a local circle over nodes in the anomaly propagation path.

[0138] In this embodiment, the method for calculating reverse coverage capability is as follows: First, for each healthy local circle, the set of nodes within that circle is extracted; then, the nodes within the circle are matched with all nodes on the anomaly propagation path, and the number of anomaly path nodes that the nodes within the circle can directly access or influence is counted. A reverse coverage score is formed by comprehensively considering weighting factors such as access latency and interaction strength. Access capability can be quantified based on indicators such as transaction response time and transaction confirmation count recorded on the chain. A higher score indicates a stronger suppression and isolation capability of the local circle on the anomaly propagation path. Through this reverse coverage measurement method, not only is the current health status of the local circle assessed, but its control capability over the anomaly path is also comprehensively judged, ensuring that the selected decision circle has effective anomaly isolation and decision execution capabilities when taking over decision-making tasks.

[0139] S3.4: Select the local blockchain circle with the largest reverse coverage capability as the decision blockchain circle. If there are multiple local blockchain circles with the same reverse coverage capability, select the local blockchain circle that is farthest from the first local blockchain circle as the decision blockchain circle.

[0140] Specifically, among all healthy local blockchain circles, the circle with the highest reverse coverage capability score is selected as the new decision-making blockchain circle. This allows for rapid isolation of the source of anomaly propagation at the system architecture level, restoring the platform's overall decision-making capability. If multiple local circles have the same reverse coverage capability score, physical location is introduced as an auxiliary screening condition to further reduce the risk of potential anomaly contagion. The physical location is calculated based on the node's geographical coordinates, tunneling location, or construction section number recorded on the chain. The circle with the furthest physical distance from the first local blockchain circle is selected to maximize physical isolation of the anomaly source and prevent potential geological or equipment anomalies from spreading along physically close paths. After selecting the decision-making circle, a decision-making circle change instruction is broadcast through the on-chain consensus mechanism, and the execution process of the anomaly isolation smart contract for generating the decision instruction of the new decision-making circle is initiated.

[0141] Please see Figure 8 The figure is a flowchart illustrating the consensus method for decision-making blockchains provided in this application embodiment. The specific steps of the method are as follows:

[0142] A1: Deploy anomaly isolation control smart contracts on the blockchain;

[0143] Specifically, in traditional remote monitoring systems, anomaly isolation typically relies on central server scheduling commands, resulting in significant response delays, incomplete coverage, and a lack of on-chain traceable isolation process records. To achieve truly rapid anomaly isolation and autonomous control within the decision-making circle, this embodiment deploys anomaly isolation control smart contracts on the overall blockchain circle. These smart contracts use on-chain evidence of construction node states, propagation path information, and selected decision-making blockchain circle information as input, automatically executing anomaly node locking, transaction freezing, and control transfer operations. Deployed on the main blockchain, initiated by decision support nodes, and officially activated after multi-node consensus confirmation, the smart contracts effectively ensure the transparency and immutability of the contract rules. Replacing traditional centralized control mechanisms with smart contracts significantly improves response speed, reduces human intervention, and enhances the overall system's autonomy and credibility.

[0144] A2: According to the path, the on-chain transaction permissions of the abnormal construction surface node and directly related nodes are initialized and locked through the abnormal isolation control smart contract;

[0145] Specifically, in a construction monitoring environment, abnormal conditions may not only affect the source node, but may also spread to adjacent nodes along the data interaction chain. If not controlled in time, it will cause abnormal data to spread on the chain and mislead the decision-making system.

[0146] In this embodiment, after the smart contract is started, it first automatically locates the abnormal construction surface node and its directly associated node set based on the abnormal propagation path recorded on the chain. Locking operations include suspending related nodes from submitting new transactions on the chain, preventing abnormal nodes from generating new blocks, and freezing the node's consensus voting rights within the local blockchain circle. Specifically, a node permission control table is set in the on-chain smart contract, marking the transaction status of the locked node as frozen, and the consensus participation list of the local blockchain circle is updated synchronously to prevent abnormal nodes from continuing to affect the on-chain state. Through this precise locking method based on the propagation path, this embodiment can effectively cut off the abnormal propagation link with minimal control, reducing the overall system load, while avoiding excessive intervention in normal nodes, thus improving the targeting and efficiency of isolation operations.

[0147] A3: Trigger the execution of anomaly isolation consensus operations by each construction site node within the decision blockchain circle, so as to confirm the scope of anomaly control based on the consensus results;

[0148] Specifically, to ensure synchronization and consistency within the decision-making circle during anomaly isolation, this embodiment designs an anomaly isolation consensus operation process. The consensus content includes, but is not limited to: confirmation of the anomaly node list, confirmation of the isolation strategy, and confirmation of recovery conditions. The consensus operation involves all active implementation nodes within the decision-making blockchain circle, employing a simplified Byzantine fault-tolerant algorithm to quickly reach a consensus. During execution, the smart contract broadcasts an isolation proposal, each node independently verifies the proposal, and submits a signed response within a specified time window; once a preset response threshold is reached, the proposal is officially approved, triggering the isolation control command to take effect. This anomaly isolation consensus mechanism ensures consistent recognition and synchronized execution of anomaly control actions within the decision-making circle, avoiding on-chain forks or decision-making chaos, and further improving the reliability and stability of anomaly management.

[0149] A4: After consensus is confirmed, the anomaly isolation control smart contract adjusts the control strategy of the abnormal construction surface node and sends the control strategy back to the decision support layer.

[0150] After the consensus operation for anomaly isolation is completed, the smart contract dynamically adjusts the control strategy for the abnormal node based on the consensus result. The control strategy includes the lock duration, unlocking conditions, and subsequent health check mechanisms. In this embodiment, the smart contract generates a health recovery monitoring plan for each locked node, periodically scheduling on-chain or off-chain monitoring modules to sample and detect changes in the abnormal node's state. If the indicators recover to the normal range after multiple consecutive checks, the frozen state is automatically lifted, and the node's transaction permissions are restored; conversely, if the anomaly persists, the freeze remains until the anomaly is resolved or manual intervention is required. All control adjustment processes are transparently recorded on-chain, generating a verifiable anomaly handling trajectory, effectively supporting audit traceability and accident liability determination during subsequent construction processes. Through this adaptive control strategy, this embodiment not only achieves timely isolation of abnormal states but also establishes a self-healing mechanism for abnormal nodes, ensuring the system's flexible recovery capability and long-term stability after anomalies occur.

[0151] The remote monitoring method further includes:

[0152] Summarize the consensus results transmitted within the aforementioned decision-making blockchain circle;

[0153] Perform a global state determination based on the consensus result;

[0154] Based on the global state, decision instructions are generated, wherein the generation of decision instructions is based on the NSGA-II algorithm, which is used to optimize objective conditions, including minimizing the escape time, minimizing resource scheduling costs, and maximizing the progress guarantee rate.

[0155] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A TBM remote monitoring method based on a multi-level collaborative management architecture, applied to a TBM remote integrated monitoring platform, wherein the platform includes a multi-level collaborative management architecture, the multi-level collaborative management architecture comprising: The system comprises a construction operation layer, a construction control layer, and a decision support layer. Each monitoring scenario's construction operation layer corresponds to multiple construction face nodes. Each monitoring scenario forms multiple local blockchain circles, each containing the corresponding multiple construction face nodes. These local blockchain circles form a unified blockchain circle, which corresponds to the decision support layer. Each decision support layer corresponds to a decision node. The TBM remote monitoring method is characterized by the following: The construction status is monitored and construction status data is collected through the construction operation layer. The construction status data is transmitted to the construction control layer, so that the construction control layer can perform construction status analysis and obtain analysis results. Multiple analysis results are aggregated into the decision support layer, so that the decision support nodes in the decision support layer can determine the decision blockchain circle for the local blockchain circle whose current state is abnormal based on the multiple analysis results, wherein the decision blockchain circle is one of the local blockchain circles whose current state is not abnormal.

2. The TBM remote monitoring method based on a multi-level collaborative management architecture according to claim 1, characterized in that, The construction control layer performs construction status analysis and obtains analysis results, including: The construction status data is standardized to obtain equipment operating parameters, geological environment parameters, and construction efficiency parameters; The equipment operating parameters are input into a preset equipment risk identification model, and the health status of the construction equipment and the risk status of the main bearing equipment are output through the equipment risk identification model. The equipment risk identification model includes a first model and a second model. The geological environment parameters and construction efficiency parameters are input into a preset geological assessment model, and the surrounding rock grade and fracture zone distribution information are output through the geological assessment model.

3. The TBM remote monitoring method based on a multi-level collaborative management architecture according to claim 2, characterized in that, The first model consists of a convolutional neural network and a long short-term memory neural network, which are used to output the health status of construction equipment. The input of the first model is the infrared temperature measurement image and vibration sequence signal in the equipment operating parameters.

4. The TBM remote monitoring method based on a multi-level collaborative management architecture according to claim 2, characterized in that, The second model is based on a multilayer perceptron network, which is used to analyze the content and viscosity changes of ferromagnetic particles in lubricating grease to predict the risk status of the main bearing equipment. The input of the second model is the oil monitoring sensor data in the equipment operating parameters.

5. The TBM remote monitoring method based on a multi-level collaborative management architecture according to claim 2, characterized in that, The core of the geological assessment model is a graph neural network, which includes a first branch and a second branch. The first branch performs a bandpass graph convolution operation, and the second branch performs a smooth graph convolution operation. The geological environment parameters and the construction efficiency parameters are input into the preset geological assessment model, including: The construction surface nodes are mapped to monitoring points, wherein the node feature vectors of the monitoring points are calculated based on the geological environment parameters and the construction efficiency parameters; Constructing an undirected graph structure based on the spatial location information of the monitoring points, wherein constructing the undirected graph structure includes: establishing a first-order edge for the monitoring point pair when the Euclidean distance between the monitoring points is less than a first distance threshold; and establishing a second-order edge for the monitoring point pair when the Euclidean distance between the monitoring points is greater than or equal to the first distance threshold and less than a second distance threshold. The edge weights of the first-order and second-order edges are calculated based on the node feature vectors of the monitoring points to construct an undirected graph structure.

6. The TBM remote monitoring method based on a multi-level collaborative management architecture according to claim 5, characterized in that, Based on the first branch, a bandpass graph convolution operation is performed on the first subgraph formed by the first-order edge to obtain the first feature matrix for lithological abrupt changes at the monitoring point. Based on the second branch, a smooth graph convolution operation is performed on the second subgraph formed by the second-order edge to obtain a second feature matrix for the continuity of the monitoring point segment. In a pre-defined multi-layer gating network, the first feature matrix and the second feature matrix are subjected to diffusion and aggregation operations in sequence according to the layer number; After each layer of operation is completed, the degree of difference of the same monitoring point in the feature matrix is ​​calculated; If the difference is greater than the preset difference threshold, then in subsequent layers, the element corresponding to the second feature matrix of the monitoring point will stop being updated, while the diffusion of the element corresponding to the first feature matrix will be retained. If the difference is less than or equal to the difference threshold, then in subsequent layers, the element corresponding to the first feature matrix of the monitoring point will be stopped from being updated, and the aggregation of the element corresponding to the second feature matrix will be retained. After all gating network operations are completed, select the corresponding elements in the feature matrix of each monitoring point that is still active to form a target feature matrix. Then, input the target feature matrix into the classification layer and output the surrounding rock grade result and the probability of fracture zone distribution of the monitoring point.

7. The TBM remote monitoring method based on a multi-level collaborative management architecture according to claim 6, characterized in that, The diffusion and polymerization operations include: When the layer number is odd, the first feature matrix is ​​locally diffused through the diffusion gating unit, and the updated first feature matrix of this layer is input to the second feature matrix through the attention mechanism as an additional input for the aggregation of the second feature matrix in the next layer. The second feature matrix is ​​the state of the matrix after the previous layer update. When the layer number is even, the second feature matrix is ​​aggregated with segment information through the aggregation gating unit, and the updated second feature matrix of this layer is input to the first feature matrix through the attention mechanism as an additional input for the diffusion of the first feature matrix in the next layer. The first feature matrix is ​​the state of the matrix after the previous layer update.

8. The TBM remote monitoring method based on a multi-level collaborative management architecture according to claim 1, characterized in that, The remote monitoring method further includes: Summarize the consensus results transmitted within the aforementioned decision-making blockchain circle; Perform a global state determination based on the consensus result; Based on the global state, decision instructions are generated, wherein the generation of decision instructions is based on the NSGA-II algorithm, which is used to optimize objective conditions, including minimizing the escape time, minimizing resource scheduling costs, and maximizing the progress guarantee rate.

9. The TBM remote monitoring method based on a multi-level collaborative management architecture according to claim 1, characterized in that, The determination of the decision blockchain circle for the local blockchain circle in the current state of anomaly includes: Mark the abnormal construction surface node where the abnormal event occurred and its corresponding first local blockchain circle; Based on the data flow relationship of all construction surface nodes maintained on the first partial blockchain circle, the propagation path of the first partial blockchain circle is deduced; Calculate the reverse coverage capability of any local blockchain circle whose current state is not abnormal along the path of propagation in the first local blockchain circle. The local blockchain circle with the largest reverse coverage capability is selected as the decision blockchain circle. If there are multiple local blockchain circles with the same reverse coverage capability, the local blockchain circle that is physically farthest from the first local blockchain circle is selected as the decision blockchain circle.

10. The TBM remote monitoring method based on a multi-level collaborative management architecture according to claim 9, characterized in that, The remote monitoring method further includes: Deploy anomaly isolation control smart contracts on-chain; According to the path, the on-chain transaction permissions of the abnormal construction surface node and directly related nodes are initialized and locked through the abnormal isolation control smart contract; Triggering each construction node within the decision blockchain circle to perform anomaly isolation consensus operation, so as to confirm the scope of anomaly control based on the consensus result; After consensus is confirmed, the anomaly isolation control smart contract adjusts the control strategy for the abnormal construction surface node.

Images